============================================================================================== RAHUL'S ML BLOG -- notes on machine learning, worked out by hand est. 2026 ============================================================================================== home | about | archive | glossary | contact ---------------------------------------------------------------------------------------------- ALL POSTS -- a book in 23 chapters (69 posts), plus references and standalone specials, every number worked by hand. First time here? Start from the beginning. CHAPTER 1 -- PREDICTING HOUSE PRICES (2026) ------------------------------------------- post 1 Part 1 -- Guessing House Prices, End to End post 2 Part 2 -- Ask the Closest Rows: Gap, Same-Ruler, and How Slow It Gets post 3 Part 3 -- The Straight-Stick Rule: Dials, Leftovers, and Numerics CHAPTER 2 -- GRADING A GUESSER (2026) ------------------------------------- post 4 Part 1 -- Two Rulers for One Guess: MSE and R^2 post 5 Part 2 -- Reading the Dials: What the Coefficients Say CHAPTER 3 -- SORTING INTO BINS (2026) ------------------------------------- post 6 Part 1 -- The S-Curve, the Four-Box Table, and Why Accuracy Lies post 7 Part 2 -- The Trade Curve: Sliding the Cutoff and What AUC Measures post 8 Part 3 -- Leash and Cloud: L2 Punishment and the Two-Cloud Wall post 9 Part 4 -- Picking Settings, Skewed Piles, and Averaging Many Classes CHAPTER 4 -- HUMBLE DIALS AND WOBBLE BANDS (2026) ------------------------------------------------- post 10 Part 1 -- The Leash: Ridge, Lasso, and Humbling the Dials post 11 Part 2 -- One Dial Is a Lie: Bootstrap, Wobble Bands, and the Free Exam post 12 Part 3 -- The Dial by Hand: Where the Dials Really Come From CHAPTER 5 -- QUESTION CHARTS AND COMMITTEES (2026) -------------------------------------------------- post 13 Part 1 -- Question Charts: Building a Tree by Hand post 14 Part 2 -- The Mixing Ruler: Gini, Information Gain, and Pruning post 15 Part 3 -- Committees: Bagging, Random Forest, and Boosting CHAPTER 6 -- FINDING PATTERNS WITHOUT ANSWERS (2026) ---------------------------------------------------- post 16 Part 1 -- Looking at a Sheet With No Answers: Means, Distance, and the Ruler Problem post 17 Part 2 -- The Strongest Direction: Crushing a Many-Wall Room Into a Flat Page (PCA) post 18 Part 3 -- Grouping by Nearest Centre: K-Means From a Blank Sheet post 19 Part 4 -- The Family Tree: Hierarchical Clustering and the Dendrogram post 20 Part 5 -- Both Tools on NCI60: PCA and Clustering on Real Gene Data post 21 Part 6 -- Filling the Blanks: Recommender Systems and Matrix Factorisation CHAPTER 7 -- BUILDING A NEURAL NETWORK FROM SCRATCH (2026) ---------------------------------------------------------- post 22 Part 1 -- Stacked Rooms and One Walk by Hand: How a Network Computes a Guess post 23 Part 2 -- Rolling Downhill by Hand: How a Network Learns CHAPTER 8 -- KEEPING A NETWORK HONEST (2026) -------------------------------------------- post 24 Five Machines Against Memorising: A Tax, a Coffee Break, a Fire Alarm, and a Humbler CHAPTER 9 -- MACHINES THAT LOOK AT PICTURES (2026) -------------------------------------------------- post 25 Part 1 -- A Magic Paper Slid Over a Photo: How a Picture Network Sees post 26 Part 2 -- The Deep Factory: Humbler, Send-Home, and the Confusion Sheet CHAPTER 10 -- MACHINES THAT READ WORDS (2026) --------------------------------------------- post 27 Part 1 -- Words Into a Machine: The Notepad and the Walking Worker post 28 Part 2 -- The Two-Memory Worker: How an LSTM Remembers Far-Back Words post 29 Part 3 -- The Look-Across Machine: Attention and the Transformer by Pencil by hand: The Walking Machine and the Vault -- RNN and LSTM by Pencil by hand: LSTM From Pencil -- RNN and LSTM From Scratch, Nothing But a Pencil CHAPTER 11 -- ATTENTION GROWS EYES (2026) ----------------------------------------- post 30 Part 1 -- The Vision Transformer: A Photo Cut Into Strips That Look at Each Other post 31 Part 2 -- Encoder or Decoder: The One Mask That Splits BERT From GPT by hand: Transformer From Pencil -- Attention From Scratch, One Number at a Time by hand: Attention and the Transformer by Pencil -- Every Word Looks at Every Word by hand: The Mark That Tells a Transformer Where It Is -- Positional Encoding Built From Scratch by hand: Transformers With Pencil -- A Whole Block Worked by Hand, One Line at a Time by hand: Vision Transformer From Pencil -- Strips, Seat-Stamps, and Masks by Hand CHAPTER 12 -- REINFORCEMENT LEARNING (2026) ------------------------------------------- intro: The Agent Ladder -- Five Cats, One New Power Each, From Reflex to Learning post 32 Part 1 -- Bandits and Exploration: The Greedy Shopkeeper and His Faulty Machines post 33 Part 2 -- Worth and Bellman from Zero: What a Spot Is Worth When the Future Branches post 34 Part 3 -- Grading a Plan by Pencil: A Table of Lies That Heals Into the Truth post 35 Part 4 -- Finding the Best Plan by Pencil: How Good Arrows Spread From the Runway post 36 Part 5 -- No Map in the Arguments: Value Iteration Built by Pencil CHAPTER 13 -- SAMPLE-BASED LEARNING (2026) ------------------------------------------ post 37 Part 1 -- No Die, Just a Sample: TD(0) Built by Pencil post 38 Part 2 -- Max or Honest Average: Q-Learning and Expected Sarsa by Pencil post 39 Part 3 -- A Notebook and a Rehearsal: Dyna-Q by Pencil post 40 Part 4 -- A Clock for Curiosity: Dyna-Q+ by Pencil ref: The Sample-based Map: MC, TD, DP, and the Sarsa Family CHAPTER 14 -- FUNCTION APPROXIMATION (2026) ------------------------------------------- post 41 Part 1 -- Too Many Spots for a Table: State Aggregation by Pencil post 42 Part 2 -- Sharp From Blurry: Tile Coding by Pencil post 43 Part 3 -- Letting the Car Choose: Sarsa Control by Pencil post 44 Part 4 -- Two Stages and a Bend: A Q-Network by Pencil CHAPTER 15 -- TRAINING THE Q-NETWORK (2026) ------------------------------------------- post 45 Part 1 -- One Miss, a Thousand Nudges: Backpropagation by Pencil post 46 Part 2 -- A Smarter Step: Adam and Replay by Pencil post 47 Part 3 -- The Frozen Twin: A Batch of Misses at Once post 48 Part 4 -- The World Calls Three Times: Wiring the Agent by Pencil post 49 Part 5 -- From Eight Dials to a Soft Landing: The Whole Agent by Pencil CHAPTER 16 -- POLICY GRADIENT (2026) ------------------------------------ post 50 Propose and Grade: Actor-Critic by Pencil CHAPTER 17 -- TWO THE SAMPLE-BASED MAP SKIPPED (2026) ----------------------------------------------------- post 51 Part 1 -- Wait for the Whole Trip: Monte Carlo by Pencil post 52 Part 2 -- Fix the Biggest Surprise First: Priority Sweeping by Pencil CHAPTER 18 -- CHOOSING THE MACHINE (2026) ------------------------------------------ post 53 Four Questions Before Any Arithmetic: Which Machine for Which World CHAPTER 19 -- LEARNING BY COPYING (2026) ----------------------------------------- post 54 Part 1 -- The Diary and the Eight Envelopes: Learning by Copying post 55 Part 2 -- The Copying Machine by Pencil post 56 Part 3 -- One Question, Two Right Answers: Where Copying Breaks post 57 Part 4 -- Teaching the Wind: Flow Matching by Pencil post 58 Part 5 -- Riding the Wind: From Noise to an Answer CHAPTER 20 -- THE PUSH-T MACHINE (2026) ---------------------------------------- post 59 Part 1 -- From Five to Twenty-Two: Why Flow Matching Needs a Clock post 60 Part 2 -- Wrong Wind Shows Every Dial Which Way to Turn post 61 Part 3 -- Eight Blind Ticks, Then The Overshoot CHAPTER 21 -- SELF, NET, AND THE CALL THAT DOES NOT LOOP (2026) ----------------------------------------------------------------- post 62 Part 1 -- One Dot Away From An Infinite Loop CHAPTER 22 -- THE FATE OF EVERY NUMBER (2026) ---------------------------------------------- post 63 Part 1 -- The 90 Comes Back As 89.94 post 64 Part 2 -- The 52 Forces The Flip post 65 Part 3 -- Born, Works, Dies -- Then All Twelve Dials Turn CHAPTER 23 -- POLICY GRADIENTS (2026) -------------------------------------- post 66 Part 1 -- No Answer Key, Only a Score post 67 Part 2 -- Why the Log Times the Score Turns the Dial post 68 Part 3 -- Subtract a Yardstick, Lose No Truth post 69 Part 4 -- Trust the Critic a Little: GAE by Pencil REFERENCES -- flip-to decoders, one per cluster of chapters ----------------------------------------------------------- ref: A. Classification Reference -- Loss, Leash, Grid, and All the Terms (companion to Chapter 3) ref: B. Distance and Clustering Reference -- Rulers, Traps, and Ethics (companion to Chapter 6) ref: C. LSTM From Pencil -- RNN and LSTM From Scratch (companion to Chapter 10) ref: D. Transformer From Pencil -- Attention From Scratch (companion to Chapter 10) ref: E. Vision Transformer From Pencil -- Strips, Seat-Stamps, and Masks (companion to Chapter 11) ref: The Sample-based Map -- MC, TD, DP, and the Sarsa Family (companion to Chapter 13) STANDALONE SPECIALS -- written for Hacker News, each self-contained ------------------------------------------------------------------ special: Genetic Algorithm From Scratch -- Optimising With No Gradient special: Cheapest Walk: UCS and A* -- Every Pop Shown, the Proof Included special: Informed Search Jargon -- One Machine, Three Slips, Every Word Debunked special: Bayes by Head-Count -- The Greedy Shopkeeper and His Faulty Machine special: The Walking Machine and the Vault -- RNN and LSTM by Pencil special: Attention and the Transformer by Pencil -- Every Word Looks at Every Word special: The Mark That Tells a Transformer Where It Is -- Positional Encoding From Scratch special: The Window That Drops Its Best Old Match -- Sliding Window Attention by Pencil special: The Two Flips That Are Not the Same Flip -- Attention's Shapes by Pencil special: Sliding Window Self-Attention, Built From One Pencil And One Page special: Build a GPT, Forced -- A Reading Ladder special: A Writing Machine of One Room -- Smallest GPT That Works, by Pencil special: What a Machine Does When You Press "h" -- Every Matrix, Forced Into Being special: What Happens When You Press "h" in a GPT -- A Plain Walkthrough special: Press "h" -- Final GPT Pencil Pass special: A Note a Machine Keeps So It Stops Redoing Old Work -- KV Cache by Pencil special: The Agent Ladder -- Five Cats, One New Power Each special: Transformers With Pencil -- A Whole Block Worked by Hand ---------------------------------------------------------------------------------------------- home . source on GitHub ==============================================================================================